fix(dual-franka): align RLinf v0.4 deployment and control diagnostics (v0.1) by NewYork424 · Pull Request #262 · RLinf/RPent

NewYork424 · 2026-10-10T17:04:41Z

Description

Align dual-Franka deployment with official RLinf v0.4 robot interfaces while retaining full-resolution observations and the existing primitive result contracts.

Use supported arm, end-effector and synchronous joint-reset interfaces. Non-gripper commands hold the grippers without reissuing open/close commands.
Keep policy images and uncropped RGB-D from the same camera acquisition. Read remote-camera intrinsics through the existing connection, reject stale projection snapshots, and follow the single-Franka camera metadata field names.
Rotate through bounded orientation references using the existing RLinf action path. After rotation, reuse move_delta to correct residual position drift beyond the existing 0.006 m position tolerance. Drift above rotation_hold.max_drift_m (default: 0.10 m) aborts the call before translation compensation. Forward optional robot.compliance settings to the official backend.
Add a continuous VLA-only diagnostic runner, tests/e2e_tests/dual_franka/dual_franka_vla_only.py, reusing the existing diagnostic validators. It connects to existing services without resetting or recovering the robot and checks Ctrl-C/STOP between individual action RPCs. A read-only vla.status returns only the prediction chunk length required by this runner.
Include Task 6 for the two-cup sorting setup, reusing Task 4 and its deployed VLA instruction.

The existing dual_franka_vla.py console retains its 20-step validation default and explicit --expected-action-steps override. Motion results retain their existing fields; rotation reports success only when compensation, if needed, succeeds and the final position and angle meet their tolerances. During posture recovery, ordinary rotation accuracy remains diagnostic; protective drift errors interrupt the recovery.

Validation

Earlier PR-related offline tests: 199 passed, 4 skipped. The final rotation, threshold/configuration and related regression tests: 35 passed.
Full unit suite: 1216 passed, 5 skipped, 3 failed. All three failures also reproduce on the unchanged base: two require the unavailable lerobot.datasets dependency, and one depends on this machine's Codex tool list.
Both configured pre-commit Ruff hooks passed over all tracked files, with no file changes. Explicit file arguments were used because the local Git version does not support the --deduplicate option used by --all-files.
Combined test additions and deletions across fix(dual-franka): align RLinf v0.4 deployment and control diagnostics (v0.1) #262–265: 1167 lines.

Both English and Chinese Sphinx builds passed with warnings treated as errors. The example configuration and both deployment guides describe possible drift sources and distinguish the 0.10 m abort threshold from the 0.006 m compensation tolerance. This update does not claim new real-robot validation of its default. A stop request takes effect after any in-flight RPC returns; it is not a hardware emergency stop.

Consolidates #253 and #254, and replaces #252. Earlier discussion remains in #252.

…r grids

…hold

…untime

…lumbing

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

Sonorous281 · 2026-10-12T07:27:24Z

+            status = model.status(timeout_s=10)
+            expected_steps = int(status["config"]["openpi"]["action_chunk"])
+            if expected_steps <= 0:
+                raise ValueError("expected action steps must be positive")


the default now requires vla.status, but both docs and tests' README.md still describe a fixed 20-step default

Sonorous281 · 2026-10-12T07:29:05Z

+def build_parser():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--skill", choices=["vla_full_rollout"], default="vla_full_rollout"


will "--skill" change execution? Do we need it?

Sonorous281 · 2026-10-12T07:29:56Z

+from robots.dual_franka.tasks import get_dual_franka_task
+from rpent.robots.components.pi05_vla_client import Pi05VLAClient
+from rpent.utils.rpc import make_rpc_client
+from tests.e2e_tests.dual_franka.dual_franka_vla import (


tests/README.md places real-robot diagnostics under tests/e2e_tests. could we put this script alongside dual_franka_vla.py, whose helpers it already uses?

qurakchin · 2026-10-12T07:24:02Z


 @pytest.mark.parametrize("embodiment", ["libero", "dual_franka", "robodojo", "yam"])
-def test_cli_forwards_model_configuration(monkeypatch, embodiment):
+@pytest.mark.parametrize("repo_id", [None, "test/dataset"])


what's this test change for?

qurakchin · 2026-10-12T07:25:57Z

        delta_xyz: Annotated[list[float], Field(min_length=3, max_length=3)],
    ) -> ToolResult:
-        """Move one Franka TCP by a bounded world-frame xyz delta in meters.
+        """Move one Franka TCP by a bounded world-frame xyz delta in meters. Completion indicates execution ended; target_reached and final_error_m report accuracy, not a mandatory retry condition.


what's the added comment means?

qurakchin · 2026-10-12T07:28:08Z

      {
        "name": "move_delta",
-        "description": "Move one Franka TCP by a bounded world-frame xyz delta in meters.",
+        "description": "Move one Franka TCP by a bounded world-frame xyz delta in meters. Completion indicates execution ended; target_reached and final_error_m report accuracy, not a mandatory retry condition.",


strange description for mcp

qurakchin · 2026-10-12T07:28:33Z

            return {
-                "ok": error <= self.controller["move_tolerance_m"],
+                # Completion describes execution, not target accuracy.
+                "ok": True,


ok is always True?

Sonorous281 · 2026-10-12T07:36:38Z

Could we tidy up the dual-Franka test directories and file names? They seem a bit scattered and inconsistent.

Sonorous281 · 2026-10-12T07:52:06Z

+            drift_limit = self.controller["rotate_max_drift_m"]
+            max_position_error = 0.0
+            exit_reason = "iteration_limit"
+            integral = np.zeros(3, dtype=np.float64)


could you explain why holding the initial position with the RLinf controller is not enough here? the position and rotation integral corrections add quite a bit of control logic. Can we simplify this while keeping the tracking fixes we actually need?

Agreed. Removed the position and rotation integral corrections and settling logic. Rotation now holds the initial TCP position and sends bounded orientation references through RLinf’s existing action path, with drift protection retained.

Sonorous281 · 2026-10-12T07:56:52Z

        }
    )
    controller = flatten_control(CONTROL)
+    rotation_hold = _require_mapping(raw.get("rotation_hold", {}), "rotation_hold")


Rotation compensation is enabled by default, but rotation_hold is missing from the example config and deployment docs. could we document the settings, how to disable compensation, and the setup used to validate the defaults?

Integral compensation and its configuration options have been removed. The remaining rotation_hold.max_drift_m guard is documented in the example config and both deployment guides, with a default of 0.03 m. This update includes offline validation, without claiming new real-robot validation of that threshold.

Code-Hit-ai · 2026-10-12T08:17:57Z

I’ve reviewed the implementation. This PR aligns the integration with RLinf, fixes the previously discussed bugs in rotation and other primitives, and adds manual VLA debugging support. These changes have practical value, and I support the functional direction. However, formatting issues still need to be addressed.

read-the-docs-community · 2026-10-12T13:20:56Z

Documentation build overview

📚 RPent | 🛠️ Build #35095484 | 📁 Comparing 1c1d393 against latest (ff91f14)

🔍 Preview build

1 file changed

± rst_source/real_world_robots/dual_franka.html

read-the-docs-community · 2026-10-12T13:20:59Z

Documentation build overview

📚 RPent-cn | 🛠️ Build #35095485 | 📁 Comparing 1c1d393 against latest (ff91f14)

🔍 Preview build

1 file changed

± rst_source/real_world_robots/dual_franka.html

qurakchin · 2026-10-12T14:39:33Z


 RPent 会将该机器人配置转换成内部双节点 RLinf cluster 和环境对象。如需使用其他文件，请传入 ``--robot-config /path/to/robot_config.yaml``。

+``rotation_hold.max_drift_m`` 默认为 ``0.03`` m。``rotate_delta`` 中，TCP 相对起始位置的漂移超过该值时，会通过现有 RPC 错误通道中止旋转。


add a simple description of the problem solved by this.

qurakchin · 2026-10-12T14:56:36Z

+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Manual-only continuous VLA rollout; never registered with the agent toolkit."""


is this file with vla-only? "vla-only" is better than "full-vla"

codex added 14 commits October 10, 2026 21:13

fix(dual-franka): rotation-hold review slice

2aae8e7

fix(dual-franka): rlinf-v04 review slice

a85d370

fix(dual-franka): manual-vla review slice

a747738

test(vla): focus manual rollout coverage on operator-visible behavior

34f70ee

test(control): use representative motion outcomes instead of paramete…

796a1f7

…r grids

test(rlinf): retain the official environment-to-worker contract only

ee1cbcf

Merge branch 'nieyi/split-01-rlinf-v04' into nieyi/split-02-rotation-…

ead897f

…hold

Merge branch 'nieyi/split-01-rlinf-v04' into nieyi/split-03-manual-vla

f5762ae

test(control): keep one full rotation control regression

face067

test(vla): retain startup and continuous policy flow coverage

f687f5b

test(control): drop deployment-specific override permutations

ae10d42

test(vla): reuse the existing diagnostic test module

99d3245

Merge branch 'nieyi/split-03-manual-vla' into nieyi/compress-franka-r…

7d8769f

…untime

refactor(dual-franka): trim unused control trace and manual rollout p…

f3e20d3

…lumbing

Copilot AI balanced review requested due to automatic review settings October 10, 2026 17:04

Copilot AI reviewed Oct 10, 2026

NewYork424 mentioned this pull request Oct 10, 2026

fix(dual-franka): align RLinf v0.4 deployment and control diagnostics #252

Closed

codex added 4 commits October 11, 2026 16:06

fix(dual-franka): propagate protective rotation abort during recovery

c0fd6a8

fix(dual-franka): preserve official dummy worker observations

5069d54

test(dual-franka): separate schema checks from prompt fixture edits

4ad43f9

test: narrow offline regressions to essential workflows

98504cb

Sonorous281 reviewed Oct 12, 2026

View reviewed changes

qurakchin requested changes Oct 12, 2026

View reviewed changes

Sonorous281 reviewed Oct 12, 2026

View reviewed changes

fix(dual-franka): restore move_delta accuracy result contract

1481f56

codex added 3 commits October 12, 2026 17:42

fix(dual-franka): restore skill results and simplify rotation

777bb10

fix(dual-franka): preserve recovery accuracy diagnostics

282dfc4

fix(dual-franka): include task 6 for default manual rollout

f814943

NewYork424 mentioned this pull request Oct 12, 2026

refactor(dual-franka): separate task policies from robot prompts #265

Open

codex added 4 commits October 12, 2026 18:04

test(dual-franka): place full VLA diagnostic under e2e

b8eed29

fix(dual-franka): finish rotation references before tolerance checks

73dcdbb

docs(dual-franka): document rotation drift guard default

dc9931b

fix(dual-franka): simplify deployment adapters and diagnostics

336ed83

NewYork424 changed the title ~~fix(dual-franka): align RLinf v0.4 deployment and control diagnostics~~ fix(dual-franka): align RLinf v0.4 deployment and control diagnostics -(v0.1) Oct 12, 2026

qurakchin reviewed Oct 12, 2026

View reviewed changes

codex added 3 commits October 12, 2026 23:11

fix(dual-franka): compensate rotation drift and document its limit

8f61c15

refactor(dual-franka): name VLA-only diagnostic and simplify docs

3b6e256

fix(dual-franka): use vla_only diagnostic name

1c1d393

Sonorous281 approved these changes Oct 12, 2026

View reviewed changes

NewYork424 changed the title ~~fix(dual-franka): align RLinf v0.4 deployment and control diagnostics -(v0.1)~~ fix(dual-franka): align RLinf v0.4 deployment and control diagnostics (v0.1) Oct 12, 2026


		RPent 会将该机器人配置转换成内部双节点 RLinf cluster 和环境对象。如需使用其他文件，请传入 ``--robot-config /path/to/robot_config.yaml``。

		``rotation_hold.max_drift_m`` 默认为 ``0.03`` m。``rotate_delta`` 中，TCP 相对起始位置的漂移超过该值时，会通过现有 RPC 错误通道中止旋转。

Conversation

NewYork424 commented Oct 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Validation

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Sonorous281 commented Oct 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Code-Hit-ai commented Oct 12, 2026

Uh oh!

read-the-docs-community Bot commented Oct 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Documentation build overview

Uh oh!

read-the-docs-community Bot commented Oct 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Documentation build overview

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

NewYork424 commented Oct 10, 2026 •

edited

Loading

read-the-docs-community Bot commented Oct 12, 2026 •

edited

Loading

read-the-docs-community Bot commented Oct 12, 2026 •

edited

Loading